feat: add NDCG metric for context ranking evaluation

[hashwnath]

Summary

Closes #230

• Adds NDCGMetric class that evaluates the ranking quality of retrieved contexts using Normalized Discounted Cumulative Gain (NDCG)
• Uses the existing context_relevancy_call LLM helper to determine binary relevance of each context item, then computes NDCG based on position ordering
• Follows the same class pattern as RetrievalPrecisionMetric (same requirements, prompt, and calculate_metric structure)

How it works

NDCG measures whether relevant contexts appear earlier in the retrieved list:

• DCG = sum(rel_i / log2(i + 1)) for each position i (1-indexed)
• IDCG = DCG of the ideal ranking (all relevant items sorted first)
• NDCG = DCG / IDCG, ranging from 0.0 to 1.0

A score of 1.0 means relevant contexts are optimally ranked. A lower score means relevant items appear later than ideal.

Edge cases handled

• Empty context list raises ValueError
• All irrelevant contexts returns 0.0 (IDCG is 0)
• Single context item works correctly

Test plan

• Unit test for compute_ndcg static method with known mathematical values (perfect ranking, worst ranking, mixed)
• Integration tests via ValidateScorer following existing parametrized test patterns (all relevant, all irrelevant, single item cases)
• test_ndcg_compute_ndcg passes locally

Generated with Claude Code